JPH04140800A - Voice recognition system using neural network - Google Patents

Abstract

PURPOSE:To enable the best recognition of a pattern which has remarkable voicing variation like a voice pattern by extracting a local feature in a layer close to the input layer of the neural network and then extracting a general feature in a high-order hidden layer. CONSTITUTION:After a phoneme pattern corresponding to a phoneme category is shown to the input layer 1, the output units in the corresponding phoneme category are set to '1' and other units are all set to '0'. The input layer 1 has a time base in the horizontal axis direction and a frequency axis in the vertical axis direction. Windowing 11 is performed locally for the time and frequency. A signal is propagated from the input layer 1 to the neural unit 21b of a 1st hidden layer 2 through a connection 11a corresponding to a nerve. In the 1st hidden layer 2, windowing 21 is performed similarly to extract a more general feature of a voice and a signal is propagated to the unit 31b in a 2nd hidden layer through a connection 21a. Similarly, signal propagation to higher-order layers is carried out. In a learning stage, an actual output value outputted by the neural network 30 is compared with the target value in the corresponding phoneme category and coupling coefficients between respective layers are so adjusted that the error becomes smaller. A recognition result is obtained as a phoneme category result having the maximum output in an output layer 5 by feature extraction in the hidden layers 2 - 4 from the input layer 1 by using the learnt coupling coefficients.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition method using a neural network, and particularly to a speech recognition method that recognizes speech patterns using a neural network.

[Prior Art and Problems to be Solved by the Invention] In speech recognition devices, it is important to realize on a computer a judgment mechanism similar to the judgment mechanism used by everyday humans, and it is important for the practical application of speech recognition. As a solution, neural networks, which are simple models of human neural networks realized on computers, are widely used.

However, conventional neural networks have a problem in that the 8-folding ability deteriorates, that is, the discrimination performance deteriorates for unlearned patterns that are different from learned patterns.

In normal neural networks, the connection from the input layer to the hidden layer and the connection from the hidden layer to the output layer are often realized by fully connected, which enables efficient feature extraction and accurate recognition of speech. This is because it cannot be said that a structure suitable for processing is necessarily realized. For this reason, a speech recognition device using a neural network can deal with temporal and single-frequency fluctuations associated with variations in the vocal style unique to speech phenomena. This was an obstacle to its realization.

Therefore, the main object of the present invention is to realize a highly accurate speech recognition method by using a network structure that is resistant to temporal and frequency variations in speech.

[Means for Solving the Problems] The present invention is a speech recognition method using a neural network, which includes a speech analysis means for analyzing continuously uttered input speech, and a time series of characteristic parameters for analyzing the analyzed speech. The neural network is configured to include a conversion means for converting into , an input layer, a plurality of hidden layers, and an output layer, and to which the converted feature parameters are input. Local windowing is performed in both the time and frequency directions, and a single unit, , , , and is configured to apply windowing to the entire layer on the input side.

[Operation] A speech recognition device using a neural network according to the present invention performs local windowing in both the time direction and the frequency direction of the input layer of the neural network and the hidden layer on the input side, and performs windowing on the input side. By performing this on the entire layer, more game-related features can be extracted in the upper hidden layer, and patterns with significant vocalization fluctuations, such as voice patterns, can be optimally recognized.

[Embodiment of the invention] FIG. 1 is a schematic block diagram of an embodiment of the invention.
FIG. 2 is a diagram showing an example of the configuration of the neural network shown in FIG. 1.

First, referring to FIG. The voice analyzed by the above is converted into a time series of characteristic parameters and is provided to the neural network 30.

As shown in FIG. 2, the neural network 30 includes an input layer 1, a hidden layer 1 2, a hidden layer 2 3, and a hidden layer 3.
It is composed of a layer 4 and an output layer 5. In the example shown in FIG. 2, the number of hidden layers is three, but generally any number of hidden layers can be used. In FIG. 2, there are six phoneme categories /b, d, g, m, n.

A neural network for identifying N/ may be described as an example, or any number and type of classification category may be selected. The neural network learning method is [
1] McCIe 11and.

J.L., D.E., Rumelhart an.
d the PDP Re5earch Gr.
.

up:'Parailel Distributed
Processing, vol. 1. , cha
p, 8. The error backpropagation method described in MIT Press (1988) is used.

First, after presenting a phoneme pattern corresponding to a phoneme category to input layer 1, the output unit of the corresponding phoneme category is set to "1", and all other units are set to "0°" (the example shown in Figure 2). In this case, the unit corresponding to /m/ is assumed to be "1").The phonetic pattern presented to the neural network 30 uses a pattern that has been subjected to FFT or LPC type feature analysis.The input layer 1 is arranged in the horizontal axis direction. represents the time axis, and the vertical axis represents the frequency axis.In the input layer 1 shown in Fig. 2, 15
Audio patterns for 16 channels are input in the vertical (frequency) direction for a frame (1,50 m5ec). Windowing 11 is performed locally with respect to time and frequency, and consists of 3 frames (30 msec) in the time direction,
Windowing is performed for four channels in the frequency direction, and a signal is propagated from the input layer 1 to the neural unit 21b of the first hidden layer 2 via a connection lla corresponding to a nerve.

Similarly, in the first hidden layer 2, windowing 21 (5 frames x 5 channels) is performed to extract more global features of the audio, and then to the unit 31b of the next hidden layer 2 3. A signal is propagated via connection 21a. Thereafter, signal propagation to upper layers is performed in the same manner. window hanger 1
1.21°31 (5 frames x 5 channels) etc. are input layer 1. Hidden layer 1st layer 2. Hidden layer 2nd layer 3° hidden layer 3rd layer 4 are continuously applied over the entire area.

In the learning stage, the actual output value output by the neural network 30 and the target value (“
1" or "O"), and adjust the coupling coefficient between each layer so that these errors are as small as possible. The recognition result is calculated from the input layer 1 to the hidden layer using the learned coupling coefficient. After the feature extraction in steps 2 to 4, a phoneme category result having the maximum output in the output layer 5 is obtained.

As described above, in a lower layer relatively close to the input layer 1, local features are extracted, and in a higher layer near the output layer 5, global features are extracted and integrated. As a result, it is possible to realize a network structure that is more resistant to changes in input patterns than the conventional case in which all the connections between layers are present.

[Effects of the Invention] As described above, according to the present invention, after local features are extracted in a layer close to the input layer of a neural network, more global features are extracted in an upper hidden layer. As a result, it is possible to optimally recognize patterns with significant vocalization fluctuations, such as voice patterns.

[Brief explanation of the drawing]

FIG. 1 is a schematic block diagram of a speech recognition system to which an embodiment of the present invention is applied. FIG. 2 is a diagram showing the configuration of the neural network shown in FIG. 1. In the figure, 1 is an input layer, 2 is a first hidden layer, 3 is a second hidden layer, 4 is a third hidden layer, 5 is an output layer, 10 is a speech analysis device, 11, 21, 31, 4 ] indicates a window, 20 indicates a conversion circuit, and 30 indicates a neural network.

Claims (1)

[Scope of Claims] Speech analysis means for analyzing continuously uttered input speech, conversion means for converting the speech analyzed by the speech analysis means into a time series of feature parameters, and an input layer and a plurality of hidden layers. and an output layer, and into which the feature parameters converted by the conversion means are input, the neural network having a local window in both the time direction and the frequency direction in the input layer and the hidden layer on the input side of the neural network. and has a single unit connection from a layer on the input side including the input layer or hidden layer to a layer on the output side including the hidden layer or the output layer, and the windowing is performed on the input side. A speech recognition method using a neural network that is characterized by performing it on the entire layer.